AI Briefing
KOSign in

OpenAI internal cyber model hits 77.0% on ExploitGym, nearly double Anthropic's Mythos 5.2

·2026.11.13 08:18

Key point

OpenAI's unreleased internal model scored 77.0% on ExploitGym, nearly double Anthropic's Claude Mythos 5.2 (39.2%) and significantly outperforming GPT-6 Cyber (52.6%).

Details

  • OpenAI's internal upcoming cyber model achieved a 77.0% success rate on the ExploitGym benchmark, a test for converting real-world vulnerabilities into working exploits across 869 containerized tasks.
  • The internal model outperformed OpenAI's released GPT-6 Cyber (52.6%) and GPT-6 Astra (42.4%), as well as Anthropic's Claude Mythos 5.2 (39.2% with a 6-hour timeout) and Claude Opus 5.5 (34.5%).
  • The 77.0% score was reached at approximately 75K output tokens, with performance hovering around 70% at 30K tokens, indicating significant headroom in OpenAI's internal capabilities compared to public releases.
  • The comparison involves different evaluation harnesses: OpenAI measures success rate across token budgets, while Anthropic uses wall-clock timeouts (2-hour and 6-hour runs) in Claude Code.
openaibenchmarkanthropiccybersecurityclaude-mythosexploitgymgpt-6
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

PyTorchKRAnthropic Releases Claude Fable 5.2 and Mythos 5.2GeekNewsAnthropic and OpenAI Replace Git Systems Due to AI Agent LoadRedditClaude Fable 5.2 designs three working hardware prototypes without circuit fixesRedditSWE-Pair Benchmark: DeepSeek V4.1 Flash Outperforms GPT-6 Luna on Dual-Bug Fixing TasksOpenAI BlogOpenAI Registers Business Ownership in ChatGPT and Launches Agent Features in BetaHacker NewsNvidia Used Circular Financing to Keep GPU Prices High, WSJ Report Claims